A novel approach for constructing and initializing a hidden layer of multilayer neural networks for pattern classification is presented. The proposed method forms hidden layer by clustering and initializing weights of hidden units using linear program...
A novel approach for constructing and initializing a hidden layer of multilayer neural networks for pattern classification is presented. The proposed method forms hidden layer by clustering and initializing weights of hidden units using linear programming so that the back-propagation algorithm can start with linear hyperplanes around input patterns rather than the whole input space. Experimental results show that the method generates the hidden layer much more efficiently and train the net faster than the ordinary back-propagation method.